ASR PRs #123

Hello @nbara :)

as mentioned in #103 (comment) I have now submitted quite a few PRs for ASR, based on a big Claude orchestrator that ran for a while, comparing the code here to published papers and other publicly available ASR code (never copying verbatim from licensed code).

I have instructed the coding agent to keep the diff minimal, and to open a dedicated PR for each issue.

I think a few PRs can be merged with clear benefits and with low-effort reviews:

• [FIX] ASR: make nonlinear_eigenspace reproducible #118

☝️ should be merged first as it is needed for other tests to pass

• [MAINT] ASR: fit_eeg_distribution docstring fixes + asr_calibrate input guards #121
• [DOC] ASR: fix numpydoc docstring rendering #119
• [MAINT] ASR: defensive hardening and docstring fixes #117
• [FIX] ASR: reuse full SHAPE_RANGE grid in clean_windows (include beta=3.5) #115
• [FIX] ASR: init clean_windows rejection masks before one-sided zthresholds #113

☝️ easy to review /clear benefit

and then the remaining PRs may need some more review / discussion. I think that can come after merging the clear ones, to clean up a little first.

• [FIX] ASR: asr_calibrate must pass step_sizes to fit_eeg_distribution #109
• [FIX] ASR: uncentered scm block covariance + overlap/bound/jump fixes #110
• [FIX] ASR: correct reconstruction dim count, blend seam, and RMS window rounding #111
• [FIX] ASR: apply online recency weights on the euclid path (weighted geometric median) #112
• [FIX] ASR: use a real PSD square root in asr_calibrate #114
• [FIX] ASR: round half away from zero for window/bad-chan counts #116
• [FIX] ASR: tolerate flat/degenerate segments in fit_eeg_distribution #120

☝️ need some closer look / review

What do you think?